Seeking Advice: Graph RAG for personal legal case

Apologies for the meandering human-authorship of my post up front; If this is the wrong sub or I should post elsewhere, I humbly thank you for correcting my mistake.

TLDR at the top for those who prefer to skip the novel.

TLDR;

I want a Graph RAG system that can represent people, places, events, statements, legal orders, governing laws, related case-law citations and the relationships between them. I'm willing to learn and build but need recommendations for existing tools, repos, guides or achievable architectures that can accomplish this without ballooning my budget.

---Novel---

I'm currently involved in a legal dispute and cannot afford to hire a lawyer. I know that representing myself puts me at serious disadvantage, but I've considered this and the math favors either giving up defending myself or doing so extremely economically.

I have roughly 20GB of discovery materials to review, mostly PDFs, emails and PNG images. The volume is beyond what I can realistically review and understand on my own so I started using Claude to make sense of it.

The size of the dataset cannot fit within a single context window so even after extracting the documents into bare text equivalents, it's still just too unwieldy.

Claude has been excellent at finding facts and details I likely wouldn't have found on my own, but to date, I have had to direct it to the specific files to get the results I need. In other words, I often need to know where to look before Claude can help me find it.

After reading and watching a fair amount of AI dev content online, I've come to the idea that I need a Graph based RAG system that can index my case data so it can be meaningfully retrieved by an LLM.

Privacy matters because this data is part of ongoing litigation (so I wouldn't publish to a public GitHub repo for example). At the same time, the data is just for my sole case and is not so sensitive that I would be worried about third parties discovering the information outside of the case itself. Given that, I need personal consumer grade privacy but not the enterprise level security and privacy one might need to build such a system for a real lawyer doing other people's legal work.

I already subscribe to Claude ($200) and Codex ($100) for my day-job, but I do not have significant additional funds to spend on this. Is there a Graph RAG setup, guide or repository that could help me combine and index these documents while capturing the semantic relationships between their contents?

If it helps thinking about it as a "user story", this is what I hope to achieve:

Given the following, Person_X said phrase_Y in a document_F about topic_Z on date_1 Person_X made statement_B about topic_Z in email_G on date_2 I want to notice that these statements relate to the same underlying topic "Z" and further that they are contradictory and given that phrase_Y was made 'under oath' in document_F, there may potentially be legal implications here worth exploring.

I want Claude to be capable of discovering these relationships or to notice and import this kind of interrelated data while I'm constructing my filings. This would help me spot lies, omissions, contradictions, changing stories, inflection points in when someone learned a fact etc.

Put in a shopping list, my goals are:

- Low maintenance costs with low cost cloud other infrastructure costs and no expensive legalAI subs

- As much offline/CPU processing as practical. I have a 10 year old gaming PC so local LLM workloads would probably be unrealistic.

- Favor procedural determinism on every practical task, saving LLM for real cognitive/decision based work. IE: Database writing probably should go through some interface/API/ MCP to ensure consistency and rule enforcement

- Right-sized for my existing claude and codex subscriptions. My work uses both, but I have some control over the need there and can leverage most of these subs for my legal work if needed.

- An index granting LLMs access to search and reason across the full dataset with efficiency

- Support for incremental updates as new filings, communications or other evidence is created

- Traceable results that point back to original documents (everything has citation/provenance chain)